geomed_copilot
<div align="center">
# RadMeasure
**A medical measurement agent with controlled verification and repair.**
HRNet landmarks · Vision-language models · KEEP / REPAIR / STOP
[Overview](#overview) · [Architecture](#architecture) · [Getting Started](#getting-started) · [Documentation](docs/DUAL_PATH_WORKFLOW.md)
</div>
## Overview
RadMeasure measures hallux valgus angle (HVA) and intermetatarsal angle (IMA) from foot radiographs. An HRNet landmark model and a vision-language model inspect the image independently; a LangGraph controller checks their agreement and decides whether to accept, recheck, or request human review.
- **Simple web interface** — upload an image and choose GPT, Claude, or Gemini.
- **Protocol-constrained actions** — model requests are limited to registered definitions and tools.
- **Bounded repair** — the VLM rechecks the image within an explicit retry budget.
- **Traceable decisions** — retain estimates, verification reasons, and review history.
## Architecture
Accepted measurements come from landmark geometry. Repair revises the VLM estimate, not the landmark coordinates. Agreement alone does not establish accuracy.
## Getting Started
```bash
pip install -e '.[api,prod,vision,inference]'
```
Follow the [setup guide](docs/DUAL_PATH_WORKFLOW.md#configuration-and-local-execution) to configure the landmark checkpoint, comparison policy, and app access key. Start the API and worker in separate terminals:
```bash
uvicorn radmeasure.api:create_app --factory --host 127.0.0.1 --port 8000
radmeasure-worker
```
Open **http://127.0.0.1:8000**, upload an image, and enter your selected provider's Model ID and API key. Checkpoints and credentials are not bundled.
---
[Workflow & configuration](docs/DUAL_PATH_WORKFLOW.md) · [Controller source](src/radmeasure/measurement_graph.py) · [Provider adapters](src/radmeasure/provider_transport.py) · [Contributing](CONTRIBUTING.md)
*Research prototype; not validated for clinical use.*
TDQS
Scored across 3 tools
Each tool serves a clearly distinct purpose: describing backend capabilities, listing available cases, and analyzing a specific radiograph. There is no overlap in functionality, making misselection unlikely.
All tool names follow a consistent verb_noun pattern: list_geomed_capabilities, list_available_cases, analyze_radiograph. The naming is uniform and predictable.
With only 3 tools, the set is tightly scoped for the stated purpose of interacting with a GeoMed backend for case analysis. Each tool earns its place, and the count is appropriate—neither too sparse nor excessive.
The tool surface covers the full expected workflow: understanding the backend's capabilities, selecting a case, and performing the analysis (which includes retrieval and citations). No additional operations seem necessary for the apparent read-only analysis domain.